Refatoração Antropológica: proposições para repensar diversidade de gênero e raça em sistemas de Inteligência Artificial

Abstract

This research is an attempt to conceive the facial analysis machine DeepFace as a code-interlocutor, an effort to expose the invisible topology of algorithmic infrastructures, that which remains repressed beneath the surface of the interface, where decision processes are inscribed in the data and in the data-to-be-seen, between bodies and their codification. Starting from the problematization of the invisibility of decision layers, the research analyzed how phenotypic and identity attributes are captured, classified, and transformed into numerical variables manipulable by facial analysis. The aim was to investigate the processes through which human subjectivities are transposed into technical-computational parameters in artificial intelligence–based facial analysis systems such as DeepFace. The research is constituted as an ethnography of science and technology within artificial intelligence laboratories, AI study groups, code analysis, participant observation in programming language workshops and hackathons, drawing on the ethnographic approach of Hess (2001) and Forsythe (1993, 2001). As a conceptual contribution, I proposed the Anthropological Refactoring (RANT), understood less as a technical remediation and more as a gesture of de-automatization of classificatory processes, which materialized in the development of the DAI (Diversity in Artificial Intelligence) algorithm. Conceived as a counterpoint to the biases of DeepFace, it aimed to establish equity in terms of gender and race within technology, but ended up specializing in colorism (Devulsky, 2021), demonstrating that even when mathematical attempts are made to balance biases, the scale tips to another side; in other words, there is no equity in the internal balance of algorithms. More than seeking definitive answers, what is offered here is a provocation, an invitation to reflect on what it means to teach a machine to see, to recognize, to classify, interrogating the worlds we are coding, the bodies we are translating into numerical vectors, and the futures that, line by line, we are consciously or unconsciously programming.